In [58]:
from IPython.display import HTML

HTML(
    """
<script
    src="https://cdnjs.cloudflare.com/ajax/libs/jquery/2.0.3/jquery.min.js ">
</script>
<script>
code_show=true;
function code_toggle() {
 if (code_show){
 $('div.jp-CodeCell > div.jp-Cell-inputWrapper').hide();
 } else {
$('div.jp-CodeCell > div.jp-Cell-inputWrapper').show();
 }
 code_show = !code_show
}
$( document ).ready(code_toggle);
</script>
<form action="javascript:code_toggle()"><input type="submit"
    value="Click here to toggle on/off the raw code."></form>
"""
)
Out[58]:
In [3]:
import fim
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt

Filigree_Banner.png

Table of Contents:¶

  • Abstract
  • Problem Statement
  • Motivation
  • Data Source
  • Methodology
  • Data Exploration
  • Data Preprocessing
  • Frequent Itemset Mining Implementation
  • Association Rules Interpretation
  • Recommended Product Bundles
  • Conclusions and Insights
  • References- Abstract
  • Problem Statement
  • Motivation
  • Data Source
  • Methodology
  • Data Exploration
  • Data Preprocessing
  • Frequent Itemset Mining Implementation
  • Results and Discussion
  • Recommendations and Limitations
  • Conclusions and Insights
  • References
  • Recommended Bundles

Abstract

This study explores the application of association rule mining techniques to analyze sales data from Silver Linings, a local business in Hyderabad, India, dedicated to the preservation and promotion of the traditional art of filigree jewelry making. Over a span of seven years, the sales data were meticulously examined to identify patterns and relationships between various jewelry items, with the aim of generating actionable insights for product bundling and customer base expansion. Employing advanced algorithms such as FP-Growth, the research delves into transactional data to unveil hidden synergies among products. The analysis distinguishes between peak and normal sales seasons, reflecting the influence of cultural festivities and everyday consumer behavior on jewelry demand. The findings underscore the significant role of seasonal trends in shaping purchasing patterns, offering a strategic framework for Silver Linings to optimize their product offerings and enhance their market position.

Problem Statement

In the competitive landscape of jewelry retail, maximizing sales and enhancing customer satisfaction are of paramount importance. Traditional bundling strategies often fail to capture the nuanced preferences of shoppers, leading to missed opportunities and suboptimal revenue generation. This project endeavors to redefine jewelry retailing by leveraging cutting-edge association rule mining techniques, such as FP-Growth and Apriori algorithms.

By analyzing transactional data from Silver Linings, the researchers aim to uncover hidden patterns and relationships between jewelry items, thereby illuminating opportunities for strategic product bundling. Their data-driven approach has the potential to revolutionize the way Silver Linings crafts its product offerings.

Employing the lens of association rule mining, the project seeks to not only boost sales but also to enhance customer satisfaction and loyalty. With these insights, Silver Linings can curate product bundles that not only aim to maximize revenue but also foster enduring relationships with its clientele.

Motivation

The motivation for this data science study is driven by a recognition of the cultural significance and the precarious state of traditional filigree jewelry-making in India. As an art form with centuries-old roots, filigree work is confronted with challenges posed by the mainstream and mass-produced jewelry and the diminishing number of skilled artisans. The preservation of such intricate craftsmanship is not only a matter of cultural heritage but also of socioeconomic importance to the communities involved in its creation.

The objective of this study is to employ data analytics to garner insights for Silver Lingings and their mission to contribute to the revitalization and sustainability of the filigree jewelry industry. It is anticipated that analyzing sales data will reveal underlying patterns that can inform strategic business decisions for them. It is posited that data-driven strategies can aid in enhancing the economic viability of artisanal crafts and ensure their continued presence in the market.

It is also envisioned that the integration of data science with traditional craftsmanship will shed light on the potential of modern methodologies to add value to cultural artifacts. This research is positioned at the intersection of cultural preservation and technological advancement, seeking to demonstrate how traditional industries can be empowered through data-driven innovation. Through this study, it is hoped that a contribution will be made not only to the survival of the filigree craft but also to the broader effort of keeping India's rich artisanal heritage alive for future generations.

Data Source

The dataset was obtained directly from the SilverLinings company owners. The initial dataset lists 11,000 item transactions across 50 columns, containing information such as customer details, billing, and shipping information. However, as will be observed later, not all features will be utilized. For the purpose of this study, the focus will be on transactional and product-related data that are pertinent to identifying purchasing patterns and behaviors.

Features irrelevant to the association rule mining, such as personal customer information, will be excluded to maintain privacy and comply with data protection regulations. The selection and preprocessing of the relevant data are critical steps that will ensure the integrity and validity of the analytical models employed. The refined dataset will enable a robust analysis of purchasing trends, revealing insights that could lead to more effective product placement, personalized marketing strategies, and optimized inventory management for SilverLinings.

In [4]:
df = pd.read_csv("sl_orders.csv", parse_dates=True, low_memory=False)
display(df.head())
display(HTML("<center><i>Table 1. Dataset Preview</i></center>"))
Name Paid at Accepts Marketing Currency Subtotal Shipping Taxes Total Discount Code Discount Amount ... Tax 4 Value Tax 5 Name Tax 5 Value Phone Receipt Number Duties Payment ID Payment Terms Name Next Payment Due At Payment References
0 #8177 2024-02-15 10:55:25 +0530 no INR 4800.0 700.0 139.81 5500.0 NaN 0.0 ... NaN NaN NaN NaN NaN NaN rv1aNObeEz15kTvgvP7lqGOpr NaN NaN rv1aNObeEz15kTvgvP7lqGOpr + x1EIiy3wXArrAcMmKx...
1 #8176 2024-02-14 13:11:57 +0530 yes INR 2520.0 0.0 73.40 2520.0 NaN 0.0 ... NaN NaN NaN NaN NaN NaN rSo9NIJNwwxPd6EXrGB9DciPl NaN NaN rSo9NIJNwwxPd6EXrGB9DciPl
2 #8175 2024-02-14 12:54:31 +0530 yes INR 1280.0 0.0 37.28 1280.0 NaN 0.0 ... NaN NaN NaN NaN NaN NaN rtEiFmdnHYy9pqg9mw4FKMLma NaN NaN rtEiFmdnHYy9pqg9mw4FKMLma
3 #8174 2024-02-14 02:48:27 +0530 no INR 8.0 0.0 0.23 8.0 NaN 0.0 ... NaN NaN NaN 9.199710e+11 NaN NaN r0uPHP4eWrBzBL0Il1ZoXCOm9 NaN NaN r0uPHP4eWrBzBL0Il1ZoXCOm9
4 #8173 2024-02-13 16:38:38 +0530 yes INR 2808.0 0.0 81.79 2808.0 NaN 0.0 ... NaN NaN NaN NaN NaN NaN rsHJLQcT8RSMspSEGQBrdGTmy NaN NaN rsHJLQcT8RSMspSEGQBrdGTmy

5 rows × 50 columns

Table 1. Dataset Preview

Methodology

Back to Table of Contents

Methodology.png

Data Exploration

Back to Table of Contents

During the examination of item names and SKUs, NaN (Not a Number) values were discovered, indicating missing information. Upon consultation with the business owners of SilverLinings, it was revealed that these discrepancies were attributable to errors in the encoding process of the SKUs. Consequently, the Lineitem name field was identified as a more reliable source for item identification in transactions.

In [5]:
df[['Name', 'Paid at',"Lineitem name",'Lineitem sku']].isna().sum()
Out[5]:
Name                0
Paid at          5406
Lineitem name       0
Lineitem sku      157
dtype: int64

Figure 1 shows the total sales count of individual products. It can be seen that Classic Floral stands out, alongside the Saree Pin and Swaying and Graceful-silver, already we can see an initial offering of perhaps products that might be bundled together given their popularity relative to sales.

In [45]:
df["Lineitem name"].value_counts()[:10][::-1].plot(
    kind="barh", edgecolor="black",
)
plt.xlabel('Amount Sold')
plt.ylabel('Product Name')
plt.text(50, -2.25, 'Figure 1. Top Selling Products');
No description has been provided for this image

In Figure 2, the sales data across time were plotted, revealing that in instances where a user purchased multiple items in a single transaction, only the data for the first item were considered, leading to the exclusion of subsequent items from being recognized as purchased on the same date. To address this issue, a forward fill method was employed, ensuring that all items in a transaction were accurately indicated as being paid for on the same date, thereby including them in the distribution analysis.

Following this adjustment, a more accurate plot of Silver Linings' sales over time was produced. Notably, a significant spike in sales was observed during the sixth, seventh, and eighth months, a pattern that was consistent across several years, with the exception of 2020. In 2020, although a similar pattern was noticeable, the volume of sales was markedly lower, which could be attributed to the impact of the Global Pandemic during that period. This analysis suggests that while seasonal trends significantly influence the sales of Silver Linings, external factors such as global events can have a profound effect on consumer behavior and overall sales performance.

In [7]:
df["Paid at"] = pd.to_datetime(df["Paid at"])
df["Paid at"] = df["Paid at"].dt.strftime("%Y-%m")
df["Paid at"].ffill(inplace=True)
fig, ax = plt.subplots()
labels = []
for i in range(2020, 2024):
    globals()[f"df_{i}"] = df.loc[
        np.where(df["Paid at"].str.contains(str(i)) == True)
    ]
    globals()[f"df_{i}"]["Paid at"].value_counts().sort_index().plot(ax=ax)

    labels.append(str(i))
ax.set_xticks(np.arange(0, 12))
ax.set_xticklabels([str(month) for month in range(1, 13)])

ax.legend(labels, title="Years")
ax.set_xlabel("Month")
ax.set_ylabel('Transaction Count')
plt.text(2.5, -130, 'Figure 2. Transactions per Month');
No description has been provided for this image

Data Preprocessing

Back to Table of Contents

It was known through conversations with the business that there were several dummy values still contained within the data, as it contained all their transactions including their test transactions when they were establishing their database. The cell below removes these dummy values to ensure the data contains only actual transactions from customers.t was known through conversations with the business that there were several dummy values still contained within the data, as it contained all their transactions including their test transactions when they were establishing their database. The cell below removes these dummy values to ensure the data contains only actual transactions from customers.

In [8]:
index_dummy = (
    df["Lineitem name"].str.extract(r"(?i)(dummy)").dropna().index.tolist()
)
index_test = (
    df["Lineitem name"].str.extract(r"(?i)(test)").dropna().index.tolist()
)
index_dummy.extend(index_test)
df.drop(index=index_dummy, inplace=True)
In [9]:
df["Lineitem name"] = df["Lineitem name"].apply(
    lambda x: x.split("(")[0].split("-")[0].strip()
)

Several products contained an extra label indicating them as variants of the main product, such as the top product ‘Classic Floral – Silver'. The purpose of removing this to simply offer a more robust offering to the customer and to the business. On the website of Silver Linings should a product have variants, customers will have the choices available to them on the page of the specific product should they examine it

In [10]:
df_678 = df.iloc[np.where(df["Paid at"].str.contains(r"\d{4}-0[678]") == True)]
df_others = df.drop(df_678.index)

Due to the observed spike in sales across the years, the dataset was split into three parts. First, the data was considered as a whole. The second part were those months that contained the spike were referred to as the peak season, while those outside of this were referred to as the regular season.

Peak Season - The period is commonly known as the Peak season, which holds significance primarily due to its association with religious observances and wedding ceremonies within the country. These occasions directly impact the demand for jewelry, as it is frequently exchanged as gifts during large family gatherings and purchased extensively for ceremonial wear. Consequently, this leads to a substantial surge in sales, contributing significantly to the business's overall performance during this period.

Normal season – During a normal season, while there may not be as many significant religious or wedding festivities, there remains a steady demand for jewelry. This demand is driven by various factors such as personal milestones, everyday celebrations, and the timeless appeal of jewelry as a form of self-expression and adornment. Although the sales may not reach the same peaks as during peak seasons, they provide a consistent flow of revenue that sustains the business throughout the year.

In [35]:
df_678["Lineitem name"].value_counts()[:10][::-1].plot(
    kind="barh", edgecolor="black"
)
plt.xlabel('Amount Sold')
plt.ylabel('Product Name');
plt.text(5, -2.5, 'Figure 3. Top Products - Peak Season');
No description has been provided for this image

The split data was examined to observe the distribution of products and their respective sales numbers. It was observed that classic floral was constant between the seasons, however during the peak season eight out of ten products were of the Silver Rakhi product line. On the other hand the standard season showed more varied products.

In [39]:
df_others["Lineitem name"].value_counts()[:10][::-1].plot(
    kind="barh", edgecolor="black"
)
plt.xlabel('Amount Sold')
plt.ylabel('Product Name');
plt.text(5, -2.5, 'Figure 4. Top Products - Regular Season');
No description has been provided for this image

The three seasons were stored in three seperate dataframes labeled: df_itemset, df_in_season and df_off_season.

In [13]:
df_itemset = df.groupby("Name")["Lineitem name"].apply(set).to_frame()
df_in_season = df_678.groupby("Name")["Lineitem name"].apply(set).to_frame()
df_off_season = (
    df_others.groupby("Name")["Lineitem name"].apply(set).to_frame()
)

Frequent Itemset Mining Implementation

Back to Table of Contents

The maximal frequency sets of the peak season were taken with the use of the pyfim package. The hyperparameters used were target="m" to return the those sets of maximal frequency, a minimum of -5 supports were set as the threshold with the negative indicating the absolute number of transactions. The zmin=2 ensures that the minimum items contained must be two.

In this example the support was set to 5 because it covered the peak season which is one quarter. However, upon consultation with Silver Linings it was communicated that their business considered a good sale to be five sales per item.

To ensure all itemsets could be viewed, the dataframe viewing options were set to ‘None’ to prevent truncation of rows and columns.

In [59]:
max_sets_in = pd.DataFrame(
    fim.fpgrowth(
        df_in_season["Lineitem name"].apply(list), 
        target="m", 
        supp=-5, 
        zmin=2
    ),
    columns=["Itemset", "Support"],
)
with pd.option_context("display.max_colwidth", None, "display.max_rows", None):
    display(HTML(f"<center>{max_sets_in.to_html()}" "<i>Table 2 - Peak Season</i></center>"))
Itemset Support
0 (Golden Rakhi Vows, Golden Bond Rakhi) 5
1 (Silver Rakhi Loyal, Silver Rakhi Suraksha) 5
2 (Rakhi Nanhe Phool, Silver Rakhi Unity) 6
3 (Rakhi Nanhe Phool, Silver Rakhi Strength) 6
4 (Bold and Beautiful, Original and Classy) 5
5 (Leaf Nazarbattu, Chakra Nazarbattu) 5
6 (Rakhi Nanhe Shaitan, Silver Rakhi Garv) 5
7 (Fine and Graceful, Swaying and Graceful) 5
8 (Swaying and Leafy, Classic Floral) 5
9 (Mesmerising Tarakasi, Original and Classy) 5
10 (Mesmerising Tarakasi, Swaying and Graceful) 5
11 (Chakra Nazarbattu, Original and Classy) 6
12 (Original and Classy, Swaying and Graceful) 6
13 (Swaying and Graceful, Classic Floral) 6
Table 2 - Peak Season

The maximal frequency sets of the standard season were taken with the use of the pyfim package. The hyperparameters used were target="m" to return the those sets of maximal frequency, a minimum of -15 supports were set as the threshold with the negative indicating the absolute number of transactions. The zmin=2 ensures that the minimum items contained must be two.

In this example the support was set to 15 because it covered the regular season which is three quarters.To ensure all itemsets could be viewed, the dataframe viewing options were set to ‘None’ to prevent truncation of rows and columns.

In [60]:
max_sets_off = pd.DataFrame(
    fim.fpgrowth(
        df_off_season["Lineitem name"].apply(list),
        target="m",
        supp=-15,
        zmin=2,
    ),
    columns=["Itemset", "Support"],
)
with pd.option_context("display.max_colwidth", None, "display.max_rows", None):
    display(
        HTML(f"<center>{max_sets_off.to_html()}" "<i>Table 3 - Regular Season</i></center>")
    )
Itemset Support
0 (Math is Fun, Swaying and Graceful) 16
1 (Fine and Graceful, Classic Floral) 21
2 (Chandbali Mini, Silver Filigree Hoops) 15
3 (Chandbali Mini, Chakra Nazarbattu) 15
4 (Petals, Original and Classy) 17
5 (Black thread Nazar Battu, Leaf Nazarbattu) 15
6 (Black thread Nazar Battu, Chakra Nazarbattu) 15
7 (Heart Nazar battu, Leaf Nazarbattu, Chakra Nazarbattu) 15
8 (Bead Dangles, Silver Filigree Hoops, Leaf Nazarbattu, Chakra Nazarbattu) 15
9 (Bead Dangles, Original and Classy) 15
10 (Bead Dangles, Swaying and Graceful) 15
11 (Silver Filigree Hoops, Swaying and Graceful) 16
12 (Leaf Nazarbattu, Chakra Nazarbattu, Original and Classy) 15
13 (Leaf Nazarbattu, Chakra Nazarbattu, Swaying and Graceful) 15
14 (Original and Classy, Classic Floral) 16
15 (Original and Classy, Swaying and Graceful) 25
16 (Classic Floral, Swaying and Graceful) 26
Table 3 - Regular Season

For the entire dataset, the support considered a four quarter year at ‘-20’ and a confidence of 75 to account for risk of the peak season but also the stability of the regular season.

In [62]:
max_sets = pd.DataFrame(
    fim.fpgrowth(
        df_itemset["Lineitem name"].apply(list), target="m", supp=-20, zmin=2
    ),
    columns=["Itemset", "Support"],
)
with pd.option_context("display.max_colwidth", None, "display.max_rows", None):
    display(HTML(f"<center>{max_sets.to_html()}" "<i>Table 4 - Full Season</i></center>"))
Itemset Support
0 (Math is Fun, Swaying and Graceful) 20
1 (Fine and Graceful, Classic Floral) 24
2 (Heart Nazar battu, Leaf Nazarbattu) 20
3 (Bead Dangles, Leaf Nazarbattu) 20
4 (Bead Dangles, Chakra Nazarbattu) 20
5 (Silver Filigree Hoops, Leaf Nazarbattu, Chakra Nazarbattu) 20
6 (Silver Filigree Hoops, Swaying and Graceful) 20
7 (Chakra Nazarbattu, Original and Classy) 22
8 (Original and Classy, Classic Floral) 20
9 (Original and Classy, Swaying and Graceful) 31
10 (Classic Floral, Swaying and Graceful) 32
Table 4 - Full Season

Fpgrowth was utilized due to its faster computation time. This is to account for the eventual growth of the dataset.

Results and Discussion

Back to Table of Contents

The dataset was segmented into three parts, corresponding to different sales seasons, a categorization established through an initial analysis of data distribution. For each season, both the minimum support thresholds and confidence levels were adjusted. This strategy was informed by discussions with the business, which indicated that selling an item five times within a quarter was deemed satisfactory. Consequently, the year was divided into four quarters, and for the peak season—which aligns with a single quarter—the minimum support threshold was established at five. The confidence level was set at 60% to mitigate the risk associated with creating bundles overly tailored to the festive season, ensuring that the product offerings remain balanced and reflective of broader consumer interests throughout the year.

Given the relatively low threshold for supports, combined with four years of sales data, a substantial number of bundles were generated. Notably, the Lift values of the top 10 bundles are exceptionally high, with the highest reaching 374.26 for the foremost bundle. This significant figure suggests that customers are 370 times more likely to purchase the items 'Ball Earrings' and 'Chakra Nazarbattu, Original and Classy' together, as opposed to purchasing them separately. This insight underscores a strong associative purchasing behavior among customers, specifically for these items, highlighting their potential as a combined offering to capitalize on during targeted sales strategies.

In season¶

In [63]:
bundles_in_season = pd.DataFrame(
    [
        ((sorted(rules[1] + tuple([rules[0]])), rules[2]))
        for rules in sorted(
            fim.fpgrowth(
                df_in_season["Lineitem name"].apply(list),
                target="r",
                supp=-5,
                conf=60,
                report="l",
            ),
            key=lambda x: (-x[2], x[0], min(x[1])),
        )
    ],
    columns=["Bundles", "Lift"],
)
bundles_in_season = bundles_in_season.loc[
    bundles_in_season.Bundles.drop_duplicates().index
].reset_index(drop=True)
with pd.option_context("display.max_colwidth", None, "display.max_rows", None):
    display(
        HTML(
            f"<center>{bundles_in_season.to_html()}" "<i>Table 5 - In Season bundles</i></center>"
        )
    )
Bundles Lift
0 [Balls Earrings, Chakra Nazarbattu, Original and Classy] 374.266667
1 [Balls Earrings, Chakra Nazarbattu, Leaf Nazarbattu] 336.840000
2 [Balls Earrings, Bead Dangles] 280.700000
3 [Balls Earrings, Silver Filigree Hoops] 280.700000
4 [Chakra Nazarbattu, Leaf Nazarbattu, Silver Filigree Hoops] 280.700000
5 [Bead Dangles, Chakra Nazarbattu, Leaf Nazarbattu] 210.525000
6 [Chakra Nazarbattu, Chakri Juda Pin, Leaf Nazarbattu] 210.525000
7 [Lotus Studs, Mesmerising Tarakasi, Original and Classy] 187.133333
8 [Balls Earrings, Mini Jhumki] 187.133333
9 [Chakra Nazarbattu, Leaf Nazarbattu, Mini Jhumki] 187.133333
10 [Balls Earrings, Leaf Nazarbattu] 140.350000
11 [Chakra Nazarbattu, Heart Nazar battu, Leaf Nazarbattu] 129.553846
12 [Balls Earrings, Black thread Nazar Battu] 99.070588
13 [Black thread Nazar Battu, Chakra Nazarbattu, Leaf Nazarbattu] 99.070588
14 [Bold and Beautiful, Original and Classy, Swaying and Traditional] 99.070588
15 [Fine and Graceful, Swaying and Graceful, Swaying and Traditional] 99.070588
16 [Balls Earrings, Chakra Nazarbattu] 86.369231
17 [Bold and Beautiful, Fine Bangles, Original and Classy] 84.210000
18 [Fine Bangles, Math is fun] 84.210000
19 [Balls Earrings, Petals] 84.210000
20 [Bold and Beautiful, Fine and Graceful, Original and Classy] 70.175000
21 [Fine and Graceful, Math is fun] 70.175000
22 [Balls Earrings, Original and Classy] 68.048485
23 [Chakra Nazarbattu, Mesmerising Tarakasi, Original and Classy] 64.776923
24 [Bold and Beautiful, Original Filigree, Original and Classy] 62.377778
25 [Classic Floral, Original Filigree, Swaying and Leafy] 62.377778
26 [Math is fun, Original Filigree] 62.377778
27 [Chakra Nazarbattu, Leaf Nazarbattu, Original and Classy] 51.036364
28 [Fine and Graceful, Original and Classy, Swaying and Graceful] 51.036364
29 [Silver Filigree Hoops, Swaying and Graceful] 45.642276
30 [Balls Earrings, Swaying and Graceful] 41.078049
31 [Chakra Nazarbattu, Leaf Nazarbattu, Swaying and Graceful] 41.078049
32 [Classic Floral, Swaying and Graceful, Swaying and Leafy] 41.078049
33 [Classic Floral, Math is fun] 20.050000
Table 5 - In Season bundles

Despite the remarkably high Lift value observed, a closer inspection of the sales volume for this top-performing bundle revealed a limitation. The minimum support threshold was established at 5, yet any attempt to exceed this threshold did not yield additional bundles, indicating that, throughout four years of collected data, these bundles were sold merely five times. Further analysis revealed that these transactions were confined to the most recent year, 2023. This temporal concentration of sales means that, according to the criteria defined by the business, these sales volumes qualify as good. This scenario highlights the nuanced nature of data analysis, where high associative strength between items (as indicated by Lift) must be contextualized with actual sales performance and temporal trends to fully understand their significance and potential for future sales strategies.

Such a finding had several possible explanations. The first was that the company was relatively young, having only started in 2017 and then having to face a global pandemic a few years later. Add to the fact that jewelry is a luxury product, and people tend not buy it in bulk. Instead, as observed by the initial data exploration, there were many sales for solo pieces like the Classic Floral line which has contributed as the top seller over the years years, that would not appear in bigger bundles usually being paired with one other item.

As for the regular season, the top bundles were closer to expectations with one of the top sellers, Original and Classy, being in the top bundle, but also the presence of other popular items as well. The support was set here at 15, and beyond such did not yield results. Which meant that at least within their quarters in the last few years, the bundles were selling well by the business's standards even when considering the pandemic's effects.

Regular season¶

In [64]:
bundles_off_season = pd.DataFrame(
    [
        ((sorted(rules[1] + tuple([rules[0]])), rules[2]))
        for rules in sorted(
            fim.fpgrowth(
                df_off_season["Lineitem name"].apply(list),
                target="r",
                supp=-15,
                conf=85,
                report="l",
            ),
            key=lambda x: (-x[2], x[0], min(x[1])),
        )
    ],
    columns=["Bundles", "Lift"],
)
bundles_off_season = bundles_off_season.loc[
    bundles_off_season.Bundles.drop_duplicates().index
].reset_index(drop=True)
with pd.option_context("display.max_colwidth", None, "display.max_rows", None):
    display(
        HTML(
            f"<center>{bundles_off_season.to_html()}" "<i>Table 6 - Regular season Bundles</i></center>"
        )
    )
Bundles Lift
0 [Chakra Nazarbattu, Chandbali Mini, Leaf Nazarbattu, Original and Classy] 208.096296
1 [Chandbali Mini, Leaf Nazarbattu, Original and Classy] 208.096296
2 [Chakra Nazarbattu, Heart Nazar battu, Swaying and Graceful] 110.168627
3 [Chakra Nazarbattu, Heart Nazar battu, Leaf Nazarbattu, Swaying and Graceful] 110.168627
4 [Chakra Nazarbattu, Heart Nazar battu, Leaf Nazarbattu, Original and Classy] 110.168627
5 [Heart Nazar battu, Leaf Nazarbattu, Swaying and Graceful] 110.168627
6 [Heart Nazar battu, Leaf Nazarbattu, Original and Classy] 110.168627
7 [Bead Dangles, Chakra Nazarbattu, Leaf Nazarbattu, Silver Filigree Hoops] 105.931373
8 [Bead Dangles, Leaf Nazarbattu, Silver Filigree Hoops] 105.931373
9 [Bead Dangles, Black thread Nazar Battu, Chakra Nazarbattu] 104.048148
10 [Bead Dangles, Black thread Nazar Battu, Leaf Nazarbattu] 104.048148
11 [Bead Dangles, Chakra Nazarbattu, Swaying and Graceful] 104.048148
12 [Bead Dangles, Chakra Nazarbattu, Leaf Nazarbattu, Swaying and Graceful] 104.048148
13 [Bead Dangles, Chakra Nazarbattu, Leaf Nazarbattu, Original and Classy] 104.048148
14 [Bead Dangles, Chakra Nazarbattu, Chandbali Mini] 104.048148
15 [Bead Dangles, Leaf Nazarbattu, Swaying and Graceful] 104.048148
16 [Bead Dangles, Leaf Nazarbattu, Original and Classy] 104.048148
17 [Black thread Nazar Battu, Chakra Nazarbattu, Silver Filigree Hoops] 89.641481
18 [Chakra Nazarbattu, Leaf Nazarbattu, Swaying and Graceful] 86.440000
19 [Chakra Nazarbattu, Heart Nazar battu, Leaf Nazarbattu] 86.440000
20 [Bead Dangles, Chakra Nazarbattu, Silver Filigree Hoops] 84.745098
21 [Black thread Nazar Battu, Leaf Nazarbattu, Silver Filigree Hoops] 83.238519
22 [Chakra Nazarbattu, Silver Filigree Hoops, Swaying and Graceful] 83.238519
23 [Chakra Nazarbattu, Leaf Nazarbattu, Silver Filigree Hoops, Swaying and Graceful] 83.238519
24 [Chakra Nazarbattu, Leaf Nazarbattu, Original and Classy, Silver Filigree Hoops] 83.238519
25 [Chakra Nazarbattu, Heart Nazar battu, Silver Filigree Hoops] 83.238519
26 [Chakra Nazarbattu, Heart Nazar battu, Leaf Nazarbattu, Silver Filigree Hoops] 83.238519
27 [Chakra Nazarbattu, Chandbali Mini, Silver Filigree Hoops] 83.238519
28 [Leaf Nazarbattu, Silver Filigree Hoops, Swaying and Graceful] 83.238519
29 [Leaf Nazarbattu, Original and Classy, Silver Filigree Hoops] 83.238519
30 [Bead Dangles, Chakra Nazarbattu, Leaf Nazarbattu] 81.355294
31 [Chakra Nazarbattu, Leaf Nazarbattu, Original and Classy] 81.037500
32 [Black thread Nazar Battu, Chakra Nazarbattu, Leaf Nazarbattu] 80.677333
33 [Chakra Nazarbattu, Chandbali Mini, Leaf Nazarbattu] 80.677333
34 [Chakra Nazarbattu, Leaf Nazarbattu, Silver Filigree Hoops] 77.341053
35 [Chandbali Mini, Leaf Nazarbattu, Silver Filigree Hoops] 74.914667
36 [Bead Dangles, Black thread Nazar Battu, Chakra Nazarbattu, Silver Filigree Hoops] 73.445752
37 [Bead Dangles, Black thread Nazar Battu, Chakra Nazarbattu, Leaf Nazarbattu, Silver Filigree Hoops] 73.445752
38 [Bead Dangles, Black thread Nazar Battu, Leaf Nazarbattu, Silver Filigree Hoops] 73.445752
39 [Bead Dangles, Chakra Nazarbattu, Original and Classy] 52.756808
40 [Chakra Nazarbattu, Original and Classy, Swaying and Graceful] 31.214444
41 [Chakra Nazarbattu, Leaf Nazarbattu, Original and Classy, Swaying and Graceful] 31.214444
42 [Chakra Nazarbattu, Heart Nazar battu, Original and Classy] 31.214444
43 [Chakra Nazarbattu, Chandbali Mini, Original and Classy] 31.214444
44 [Leaf Nazarbattu, Original and Classy, Swaying and Graceful] 31.214444
45 [Chandbali Mini, Silver Filigree Hoops, Swaying and Graceful] 26.193939
Table 6 - Regular season Bundles

Upon extending the examination to the dataset as a whole, notable patterns emerged, particularly with 'Leaf Nazarbattu' and 'Silver Filigree Hoops' frequently appearing as the second and third items in a bundle. This suggests a complementary relationship to the initially purchased item. To accommodate this analysis, the support was adjusted to -20, aiming to encapsulate the sales data across the last four years, divided into four quarters, while the confidence level was set at 75% to ensure the identification of consistently performing bundles over time.

The observation of high Lift values for these bundles might be attributed to the novelty of the products. Discussions with the business and a subsequent review of the database highlighted potential issues, likely stemming from the reliance on lineitem names over SKUs for product identification. Changes in product offerings or names, especially those not documented before the pandemic, may have introduced inconsistencies in the dataset.

Despite these challenges, the analysis unveiled valuable insights, particularly for the period following the pandemic. This is evidenced by the spike in sales highlighted in the exploratory section of the study, suggesting that recent product strategies and offerings are resonating well with customers. The findings underscore the importance of continuous data verification and the need to adapt analysis techniques to evolving product lines and market conditions, ensuring that the insights derived remain relevant and actionable.

All year round¶

In [65]:
bundles_all_year = pd.DataFrame(
    [
        ((sorted(rules[1] + tuple([rules[0]])), rules[2]))
        for rules in sorted(
            fim.fpgrowth(
                df_itemset["Lineitem name"].apply(list),
                target="r",
                supp=-20,
                conf=75,
                report="l",
            ),
            key=lambda x: (-x[2], x[0], min(x[1])),
        )
    ],
    columns=["Bundles", "Lift"],
)
bundles_all_year = bundles_all_year.loc[
    bundles_all_year.Bundles.drop_duplicates().index
].reset_index(drop=True)
with pd.option_context("display.max_colwidth", None, "display.max_rows", None):
    display(
        HTML(f"<center>{bundles_all_year.to_html()}" "<i>Table 7 - All Year Bundles</i></center>")
    )
Bundles Lift
0 [Bead Dangles, Chakra Nazarbattu, Leaf Nazarbattu, Silver Filigree Hoops] 145.820455
1 [Bead Dangles, Leaf Nazarbattu, Silver Filigree Hoops] 145.820455
2 [Bead Dangles, Chakra Nazarbattu, Silver Filigree Hoops] 132.564050
3 [Bead Dangles, Silver Filigree Hoops, Swaying and Graceful] 121.517045
4 [Chakra Nazarbattu, Heart Nazar battu, Leaf Nazarbattu, Silver Filigree Hoops] 113.760638
5 [Heart Nazar battu, Leaf Nazarbattu, Silver Filigree Hoops] 113.760638
6 [Bead Dangles, Chakra Nazarbattu, Leaf Nazarbattu] 109.234677
7 [Chandbali Mini, Silver Filigree Hoops] 104.838235
8 [Chakra Nazarbattu, Leaf Nazarbattu, Silver Filigree Hoops] 104.530792
9 [Chakra Nazarbattu, Leaf Nazarbattu, Original and Classy] 94.077713
10 [Leaf Nazarbattu, Silver Filigree Hoops, Swaying and Graceful] 91.987097
11 [Bead Dangles, Black thread Nazar Battu, Chakra Nazarbattu] 78.628676
12 [Bead Dangles, Black thread Nazar Battu, Leaf Nazarbattu] 78.628676
13 [Black thread Nazar Battu, Chakra Nazarbattu, Leaf Nazarbattu, Silver Filigree Hoops] 78.628676
14 [Black thread Nazar Battu, Leaf Nazarbattu, Silver Filigree Hoops] 78.628676
15 [Chakra Nazarbattu, Heart Nazar battu, Leaf Nazarbattu] 66.145361
16 [Chakra Nazarbattu, Silver Filigree Hoops, Swaying and Graceful] 58.795876
17 [Chakra Nazarbattu, Chandbali Mini] 55.121134
18 [Bead Dangles, Chakra Nazarbattu, Original and Classy] 37.275817
19 [Bead Dangles, Leaf Nazarbattu, Original and Classy] 37.275817
20 [Chakra Nazarbattu, Leaf Nazarbattu, Original and Classy, Silver Filigree Hoops] 37.275817
21 [Leaf Nazarbattu, Original and Classy, Silver Filigree Hoops] 37.275817
22 [Heart Nazar battu, Leaf Nazarbattu, Original and Classy] 34.946078
23 [Bead Dangles, Chakra Nazarbattu, Swaying and Graceful] 30.995652
24 [Bead Dangles, Leaf Nazarbattu, Swaying and Graceful] 30.995652
25 [Chakra Nazarbattu, Leaf Nazarbattu, Silver Filigree Hoops, Swaying and Graceful] 30.995652
26 [Heart Nazar battu, Leaf Nazarbattu, Swaying and Graceful] 29.058424
Table 7 - All Year Bundles

Recommendations and Limitations

Back to Table of Contents

Recommendations for Future Research:
Incorporate Product Variations: The current approach treated the same product with different variations (e.g., Classic Floral – oxidized vs. Classic Floral only) as identical. To refine the analysis, it is suggested that each variation be treated as a distinct product. This differentiation will allow for a more precise identification of which specific variation enjoys higher popularity, potentially influenced by factors such as production difficulty and consumer preferences.

Recommendations for the Business:
Data Cleaning on Product Records: One of the significant challenges encountered was the difficulty in distinguishing between products with similar or missing SKUs. SKUs represent a critical level of product identification, crucial for effective inventory management and sales analysis. Addressing issues related to inconsistent or missing SKUs is essential for ensuring the accuracy of product tracking, which, in turn, enhances the reliability of sales and inventory data. Implementing a more rigorous system for SKU management will contribute to the long-term efficiency and scalability of product analysis efforts.

Promoting Product Bundles: Based on the insights garnered from the study, it is recommended that Silver Linings actively promote the identified product bundles to capitalize on their sales potential. The promotion of these bundles could be strategically aligned with seasonal trends to maximize sales opportunities. By highlighting specific bundles at times when they are most likely to appeal to consumers, Silver Linings can not only increase sales but also enhance customer satisfaction through carefully curated product offerings.

Implementing these recommendations will not only refine the research approach for future studies but also provide Silver Linings with actionable strategies to continue growing their business and maintaining a competitive edge in the marketplace.

Conclusion and Insights

Back to Table of Contents

The investigation into Silver Linings’ sales data over several years has yielded profound insights into the purchasing behavior of their clientele, particularly highlighting the impact of seasonality on sales. The study successfully applied association rule mining to discern distinct purchasing patterns across peak and normal seasons, revealing that certain products, notably the Classic Floral line, consistently appealed to customers, whereas others, such as the Silver Rakhi product line, surged in popularity during peak seasons. This seasonal variance underscores the necessity for Silver Linings to tailor their marketing and product bundling strategies accordingly.

Notably, the research identified high-lift item sets that were significantly more likely to be purchased together, although these sets were predominantly found in the peak season data. This suggests that while there are potent opportunities for bundling to enhance sales, such strategies might be particularly effective during specific periods of heightened demand.

However, the study also encountered limitations due to data quality issues, such as errors in SKU encoding and the presence of dummy values, which necessitated careful data cleaning and preprocessing. This highlights the critical importance of maintaining high-quality, detailed transaction records for future analytics endeavors.

Based on these findings, several recommendations for both research and business practices were formulated. For future research, it is advised to further disaggregate product variations to precisely pinpoint consumer preferences. For Silver Linings, enhancing SKU data quality and actively promoting discovered high-lift product bundles during relevant seasons could significantly boost sales and customer satisfaction. The study not only contributes to the academic field of data analytics in retail but also provides Silver Linings with a data-driven strategy to sustain and grow their traditional filigree jewelry business amidst changing market dynamics.

Acknowledgments


The study would not have been completed if not for the support and data provided by Silver Linings.

The team would also like to acknowledge our mentor, whom without this paper would not have been possible, Prof. Christian Alis, Ph.D.

This writing of this study was assisted by the use of Chat-GPT4.

Back to Table of Contents

17.png 44.png 45.png